2 * Copyright 2015 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
24 #ifndef _DRM_GPU_SCHEDULER_H_
25 #define _DRM_GPU_SCHEDULER_H_
27 #include <drm/spsc_queue.h>
28 #include <linux/dma-fence.h>
29 #include <linux/completion.h>
30 #include <linux/xarray.h>
31 #include <linux/irq_work.h>
33 #define MAX_WAIT_SCHED_ENTITY_Q_EMPTY msecs_to_jiffies(1000)
35 struct drm_gem_object;
37 struct drm_gpu_scheduler;
40 /* These are often used as an (initial) index
41 * to an array, and as such should start at 0.
43 enum drm_sched_priority {
44 DRM_SCHED_PRIORITY_MIN,
45 DRM_SCHED_PRIORITY_NORMAL,
46 DRM_SCHED_PRIORITY_HIGH,
47 DRM_SCHED_PRIORITY_KERNEL,
49 DRM_SCHED_PRIORITY_COUNT,
50 DRM_SCHED_PRIORITY_UNSET = -2
54 * struct drm_sched_entity - A wrapper around a job queue (typically
55 * attached to the DRM file_priv).
57 * Entities will emit jobs in order to their corresponding hardware
58 * ring, and the scheduler will alternate between entities based on
61 struct drm_sched_entity {
65 * Used to append this struct to the list of entities in the runqueue
66 * @rq under &drm_sched_rq.entities.
68 * Protected by &drm_sched_rq.lock of @rq.
70 struct list_head list;
75 * Runqueue on which this entity is currently scheduled.
77 * FIXME: Locking is very unclear for this. Writers are protected by
78 * @rq_lock, but readers are generally lockless and seem to just race
79 * with not even a READ_ONCE.
81 struct drm_sched_rq *rq;
86 * A list of schedulers (struct drm_gpu_scheduler). Jobs from this entity can
87 * be scheduled on any scheduler on this list.
89 * This can be modified by calling drm_sched_entity_modify_sched().
90 * Locking is entirely up to the driver, see the above function for more
93 * This will be set to NULL if &num_sched_list equals 1 and @rq has been
96 * FIXME: This means priority changes through
97 * drm_sched_entity_set_priority() will be lost henceforth in this case.
99 struct drm_gpu_scheduler **sched_list;
104 * Number of drm_gpu_schedulers in the @sched_list.
106 unsigned int num_sched_list;
111 * Priority of the entity. This can be modified by calling
112 * drm_sched_entity_set_priority(). Protected by &rq_lock.
114 enum drm_sched_priority priority;
119 * Lock to modify the runqueue to which this entity belongs.
124 * @job_queue: the list of jobs of this entity.
126 struct spsc_queue job_queue;
131 * A linearly increasing seqno incremented with each new
132 * &drm_sched_fence which is part of the entity.
134 * FIXME: Callers of drm_sched_job_arm() need to ensure correct locking,
135 * this doesn't need to be atomic.
142 * A unique context for all the fences which belong to this entity. The
143 * &drm_sched_fence.scheduled uses the fence_context but
144 * &drm_sched_fence.finished uses fence_context + 1.
146 uint64_t fence_context;
151 * The dependency fence of the job which is on the top of the job queue.
153 struct dma_fence *dependency;
158 * Callback for the dependency fence above.
160 struct dma_fence_cb cb;
165 * Points to entities' guilty.
172 * Points to the finished fence of the last scheduled job. Only written
173 * by the scheduler thread, can be accessed locklessly from
174 * drm_sched_job_arm() iff the queue is empty.
176 struct dma_fence *last_scheduled;
179 * @last_user: last group leader pushing a job into the entity.
181 struct task_struct *last_user;
186 * Marks the enity as removed from rq and destined for
187 * termination. This is set by calling drm_sched_entity_flush() and by
195 * Signals when entity is not in use, used to sequence entity cleanup in
196 * drm_sched_entity_fini().
198 struct completion entity_idle;
202 * struct drm_sched_rq - queue of entities to be scheduled.
204 * @lock: to modify the entities list.
205 * @sched: the scheduler to which this rq belongs to.
206 * @entities: list of the entities to be scheduled.
207 * @current_entity: the entity which is to be scheduled.
209 * Run queue is a set of entities scheduling command submissions for
210 * one specific ring. It implements the scheduling policy that selects
211 * the next entity to emit commands from.
213 struct drm_sched_rq {
215 struct drm_gpu_scheduler *sched;
216 struct list_head entities;
217 struct drm_sched_entity *current_entity;
221 * struct drm_sched_fence - fences corresponding to the scheduling of a job.
223 struct drm_sched_fence {
225 * @scheduled: this fence is what will be signaled by the scheduler
226 * when the job is scheduled.
228 struct dma_fence scheduled;
231 * @finished: this fence is what will be signaled by the scheduler
232 * when the job is completed.
234 * When setting up an out fence for the job, you should use
235 * this, since it's available immediately upon
236 * drm_sched_job_init(), and the fence returned by the driver
237 * from run_job() won't be created until the dependencies have
240 struct dma_fence finished;
243 * @parent: the fence returned by &drm_sched_backend_ops.run_job
244 * when scheduling the job on hardware. We signal the
245 * &drm_sched_fence.finished fence once parent is signalled.
247 struct dma_fence *parent;
249 * @sched: the scheduler instance to which the job having this struct
252 struct drm_gpu_scheduler *sched;
254 * @lock: the lock used by the scheduled and the finished fences.
258 * @owner: job owner for debugging
263 struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f);
266 * struct drm_sched_job - A job to be run by an entity.
268 * @queue_node: used to append this struct to the queue of jobs in an entity.
269 * @list: a job participates in a "pending" and "done" lists.
270 * @sched: the scheduler instance on which this job is scheduled.
271 * @s_fence: contains the fences for the scheduling of job.
272 * @finish_cb: the callback for the finished fence.
273 * @id: a unique id assigned to each job scheduled on the scheduler.
274 * @karma: increment on every hang caused by this job. If this exceeds the hang
275 * limit of the scheduler then the job is marked guilty and will not
276 * be scheduled further.
277 * @s_priority: the priority of the job.
278 * @entity: the entity to which this job belongs.
279 * @cb: the callback for the parent fence in s_fence.
281 * A job is created by the driver using drm_sched_job_init(), and
282 * should call drm_sched_entity_push_job() once it wants the scheduler
283 * to schedule the job.
285 struct drm_sched_job {
286 struct spsc_node queue_node;
287 struct list_head list;
288 struct drm_gpu_scheduler *sched;
289 struct drm_sched_fence *s_fence;
292 * work is used only after finish_cb has been used and will not be
296 struct dma_fence_cb finish_cb;
297 struct irq_work work;
302 enum drm_sched_priority s_priority;
303 struct drm_sched_entity *entity;
304 struct dma_fence_cb cb;
308 * Contains the dependencies as struct dma_fence for this job, see
309 * drm_sched_job_add_dependency() and
310 * drm_sched_job_add_implicit_dependencies().
312 struct xarray dependencies;
314 /** @last_dependency: tracks @dependencies as they signal */
315 unsigned long last_dependency;
318 static inline bool drm_sched_invalidate_job(struct drm_sched_job *s_job,
321 return s_job && atomic_inc_return(&s_job->karma) > threshold;
324 enum drm_gpu_sched_stat {
325 DRM_GPU_SCHED_STAT_NONE, /* Reserve 0 */
326 DRM_GPU_SCHED_STAT_NOMINAL,
327 DRM_GPU_SCHED_STAT_ENODEV,
331 * struct drm_sched_backend_ops
333 * Define the backend operations called by the scheduler,
334 * these functions should be implemented in driver side.
336 struct drm_sched_backend_ops {
340 * Called when the scheduler is considering scheduling this job next, to
341 * get another struct dma_fence for this job to block on. Once it
342 * returns NULL, run_job() may be called.
344 * If a driver exclusively uses drm_sched_job_add_dependency() and
345 * drm_sched_job_add_implicit_dependencies() this can be ommitted and
348 struct dma_fence *(*dependency)(struct drm_sched_job *sched_job,
349 struct drm_sched_entity *s_entity);
352 * @run_job: Called to execute the job once all of the dependencies
353 * have been resolved. This may be called multiple times, if
354 * timedout_job() has happened and drm_sched_job_recovery()
355 * decides to try it again.
357 struct dma_fence *(*run_job)(struct drm_sched_job *sched_job);
360 * @timedout_job: Called when a job has taken too long to execute,
361 * to trigger GPU recovery.
363 * This method is called in a workqueue context.
365 * Drivers typically issue a reset to recover from GPU hangs, and this
366 * procedure usually follows the following workflow:
368 * 1. Stop the scheduler using drm_sched_stop(). This will park the
369 * scheduler thread and cancel the timeout work, guaranteeing that
370 * nothing is queued while we reset the hardware queue
371 * 2. Try to gracefully stop non-faulty jobs (optional)
372 * 3. Issue a GPU reset (driver-specific)
373 * 4. Re-submit jobs using drm_sched_resubmit_jobs()
374 * 5. Restart the scheduler using drm_sched_start(). At that point, new
375 * jobs can be queued, and the scheduler thread is unblocked
377 * Note that some GPUs have distinct hardware queues but need to reset
378 * the GPU globally, which requires extra synchronization between the
379 * timeout handler of the different &drm_gpu_scheduler. One way to
380 * achieve this synchronization is to create an ordered workqueue
381 * (using alloc_ordered_workqueue()) at the driver level, and pass this
382 * queue to drm_sched_init(), to guarantee that timeout handlers are
383 * executed sequentially. The above workflow needs to be slightly
384 * adjusted in that case:
386 * 1. Stop all schedulers impacted by the reset using drm_sched_stop()
387 * 2. Try to gracefully stop non-faulty jobs on all queues impacted by
388 * the reset (optional)
389 * 3. Issue a GPU reset on all faulty queues (driver-specific)
390 * 4. Re-submit jobs on all schedulers impacted by the reset using
391 * drm_sched_resubmit_jobs()
392 * 5. Restart all schedulers that were stopped in step #1 using
395 * Return DRM_GPU_SCHED_STAT_NOMINAL, when all is normal,
396 * and the underlying driver has started or completed recovery.
398 * Return DRM_GPU_SCHED_STAT_ENODEV, if the device is no longer
399 * available, i.e. has been unplugged.
401 enum drm_gpu_sched_stat (*timedout_job)(struct drm_sched_job *sched_job);
404 * @free_job: Called once the job's finished fence has been signaled
405 * and it's time to clean it up.
407 void (*free_job)(struct drm_sched_job *sched_job);
411 * struct drm_gpu_scheduler
413 * @ops: backend operations provided by the driver.
414 * @hw_submission_limit: the max size of the hardware queue.
415 * @timeout: the time after which a job is removed from the scheduler.
416 * @name: name of the ring for which this scheduler is being used.
417 * @sched_rq: priority wise array of run queues.
418 * @wake_up_worker: the wait queue on which the scheduler sleeps until a job
419 * is ready to be scheduled.
420 * @job_scheduled: once @drm_sched_entity_do_release is called the scheduler
421 * waits on this wait queue until all the scheduled jobs are
423 * @hw_rq_count: the number of jobs currently in the hardware queue.
424 * @job_id_count: used to assign unique id to the each job.
425 * @timeout_wq: workqueue used to queue @work_tdr
426 * @work_tdr: schedules a delayed call to @drm_sched_job_timedout after the
427 * timeout interval is over.
428 * @thread: the kthread on which the scheduler which run.
429 * @pending_list: the list of jobs which are currently in the job queue.
430 * @job_list_lock: lock to protect the pending_list.
431 * @hang_limit: once the hangs by a job crosses this limit then it is marked
432 * guilty and it will no longer be considered for scheduling.
433 * @score: score to help loadbalancer pick a idle sched
434 * @_score: score used when the driver doesn't provide one
435 * @ready: marks if the underlying HW is ready to work
436 * @free_guilty: A hit to time out handler to free the guilty job.
438 * One scheduler is implemented for each hardware ring.
440 struct drm_gpu_scheduler {
441 const struct drm_sched_backend_ops *ops;
442 uint32_t hw_submission_limit;
445 struct drm_sched_rq sched_rq[DRM_SCHED_PRIORITY_COUNT];
446 wait_queue_head_t wake_up_worker;
447 wait_queue_head_t job_scheduled;
448 atomic_t hw_rq_count;
449 atomic64_t job_id_count;
450 struct workqueue_struct *timeout_wq;
451 struct delayed_work work_tdr;
452 struct task_struct *thread;
453 struct list_head pending_list;
454 spinlock_t job_list_lock;
463 int drm_sched_init(struct drm_gpu_scheduler *sched,
464 const struct drm_sched_backend_ops *ops,
465 uint32_t hw_submission, unsigned hang_limit,
466 long timeout, struct workqueue_struct *timeout_wq,
467 atomic_t *score, const char *name, struct device *dev);
469 void drm_sched_fini(struct drm_gpu_scheduler *sched);
470 int drm_sched_job_init(struct drm_sched_job *job,
471 struct drm_sched_entity *entity,
473 void drm_sched_job_arm(struct drm_sched_job *job);
474 int drm_sched_job_add_dependency(struct drm_sched_job *job,
475 struct dma_fence *fence);
476 int drm_sched_job_add_implicit_dependencies(struct drm_sched_job *job,
477 struct drm_gem_object *obj,
481 void drm_sched_entity_modify_sched(struct drm_sched_entity *entity,
482 struct drm_gpu_scheduler **sched_list,
483 unsigned int num_sched_list);
485 void drm_sched_job_cleanup(struct drm_sched_job *job);
486 void drm_sched_wakeup(struct drm_gpu_scheduler *sched);
487 void drm_sched_stop(struct drm_gpu_scheduler *sched, struct drm_sched_job *bad);
488 void drm_sched_start(struct drm_gpu_scheduler *sched, bool full_recovery);
489 void drm_sched_resubmit_jobs(struct drm_gpu_scheduler *sched);
490 void drm_sched_resubmit_jobs_ext(struct drm_gpu_scheduler *sched, int max);
491 void drm_sched_increase_karma(struct drm_sched_job *bad);
492 void drm_sched_reset_karma(struct drm_sched_job *bad);
493 void drm_sched_increase_karma_ext(struct drm_sched_job *bad, int type);
494 bool drm_sched_dependency_optimized(struct dma_fence* fence,
495 struct drm_sched_entity *entity);
496 void drm_sched_fault(struct drm_gpu_scheduler *sched);
497 void drm_sched_job_kickout(struct drm_sched_job *s_job);
499 void drm_sched_rq_add_entity(struct drm_sched_rq *rq,
500 struct drm_sched_entity *entity);
501 void drm_sched_rq_remove_entity(struct drm_sched_rq *rq,
502 struct drm_sched_entity *entity);
504 int drm_sched_entity_init(struct drm_sched_entity *entity,
505 enum drm_sched_priority priority,
506 struct drm_gpu_scheduler **sched_list,
507 unsigned int num_sched_list,
509 long drm_sched_entity_flush(struct drm_sched_entity *entity, long timeout);
510 void drm_sched_entity_fini(struct drm_sched_entity *entity);
511 void drm_sched_entity_destroy(struct drm_sched_entity *entity);
512 void drm_sched_entity_select_rq(struct drm_sched_entity *entity);
513 struct drm_sched_job *drm_sched_entity_pop_job(struct drm_sched_entity *entity);
514 void drm_sched_entity_push_job(struct drm_sched_job *sched_job);
515 void drm_sched_entity_set_priority(struct drm_sched_entity *entity,
516 enum drm_sched_priority priority);
517 bool drm_sched_entity_is_ready(struct drm_sched_entity *entity);
519 struct drm_sched_fence *drm_sched_fence_alloc(
520 struct drm_sched_entity *s_entity, void *owner);
521 void drm_sched_fence_init(struct drm_sched_fence *fence,
522 struct drm_sched_entity *entity);
523 void drm_sched_fence_free(struct drm_sched_fence *fence);
525 void drm_sched_fence_scheduled(struct drm_sched_fence *fence);
526 void drm_sched_fence_finished(struct drm_sched_fence *fence);
528 unsigned long drm_sched_suspend_timeout(struct drm_gpu_scheduler *sched);
529 void drm_sched_resume_timeout(struct drm_gpu_scheduler *sched,
530 unsigned long remaining);
531 struct drm_gpu_scheduler *
532 drm_sched_pick_best(struct drm_gpu_scheduler **sched_list,
533 unsigned int num_sched_list);